🔍 Read the full analysis: What Evidence Can Tell Us About Learning To Code With AI on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Evidence available in 2026 shows AI can write a large share of code in some workplaces, but measured productivity gains vary and remain uncertain. Job-posting data show a thinner entry-level market, while a learning experiment found junior developers using AI scored worse on a follow-up quiz than peers who coded by hand.
AI is writing much of the code at some companies, but available evidence does not show that learning to program has become pointless. A review of 2026 research and labor-market data finds mixed results on productivity, fewer entry-level software postings, and an experiment in which junior developers who learned with AI performed worse on a follow-up quiz than those who coded by hand.
Anthropic reported that Claude authored more than 80% of code merged into the company’s codebase in May 2026, compared with “the low single digits” before Claude Code launched in February 2025. It also said the typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024. Those are company-reported measures, not an independent industry-wide estimate. Anthropic cautioned that lines of code measure quantity rather than quality and that the eightfold figure “is almost certainly an overstatement” of the productivity gain.
Controlled tests do not produce a simple answer about speed. In a 2025 randomized trial, the nonprofit METR studied 16 experienced open-source developers completing 246 tasks in projects they already knew. They took 19% longer when allowed to use the AI tools tested, although they believed the tools had made them 20% faster. A separate 2022 test of 95 developers writing a small JavaScript server found that participants using GitHub Copilot finished 55.8% faster. The studies involved different tasks, tools and settings.
Labor-market indicators also point in different directions. Indeed data cited in the review put entry-level roles at 4.5% of US software development postings in the first quarter of 2026, while senior roles made up 69.3%. Stanford researchers reported that employment for 22- to 25-year-olds in AI-exposed jobs was 19% below where it otherwise would have been. But Indeed’s US software development postings were about 23% higher year over year in November 2026, after reaching a low in May 2025. These figures describe different measures and do not, by themselves, establish that AI caused hiring changes.
Coding Skills Beyond Code Generation
For people choosing what to study, the evidence supports a distinction between typing code and understanding software. AI can generate code, but developers still need to define the task, fit new work into existing systems, test results and identify security problems. Those responsibilities matter when generated code is incomplete, incorrect or poorly suited to a project.
The learning findings add a separate concern: using a tool to finish a task may not teach the underlying concepts as effectively as doing the work. In an Anthropic experiment, mostly junior developers learning a new library scored 50% on a follow-up quiz, compared with 67% among participants who coded by hand. The gap was largest in debugging. The result is a warning about one learning setup, not proof that every use of AI impairs learning.
For employers, students and career changers, the stakes are practical. If junior hiring remains limited, newcomers may have fewer opportunities to build experience through routine coding tasks. At the same time, available figures do not show that demand for developers has disappeared or establish which skills employers will prioritize across the industry.
Products Worth Considering

FLEXISPOT 48 x 24-Inch Solid Rubberwood Electric Standing Desk, Black Frame

Samsung 32-inch ViewFinity S8 4K UHD IPS Monitor with USB-C and KVM

Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 256GB SSD, Liquid Retina Display, and Blush Finish
As an affiliate, we earn on qualifying purchases.
From Code Output to Work
The debate was sharpened by a prediction from Anthropic chief executive Dario Amodei. Speaking at the Council on Foreign Relations on March 10, 2025, he said AI would be “writing 90 percent of the code” in three to six months and “essentially all of the code” in twelve. Anthropic’s later account of its own codebase suggests extensive adoption at the company, though the figures do not independently verify the prediction across software businesses.
Amodei’s qualification was that programmers still need to specify what is being built, how it fits with other code and whether the design is secure. That distinction helps explain why code volume alone is not a measure of productivity. A generated line must still be reviewed, tested and maintained, and the value of faster drafting depends on whether the broader task is completed well.
Other company estimates are not directly comparable. Google chief executive Sundar Pichai said in October 2024 that AI generated more than a quarter of new Google code, which engineers then reviewed and accepted. METR cited an estimate in February 2026 that roughly 4% of GitHub commits were authored by Claude Code. The figures cover different companies, dates and measures, so they should not be treated as a single industry-wide adoption rate.
“The programmer still needs to specify what is being built, how it fits with other code and whether this is a secure design or an insecure design.”
— Dario Amodei, Anthropic chief executive, speaking at the Council on Foreign Relations on March 10, 2025
Limits of the Available Evidence
How broadly the findings apply remains uncertain. Anthropic’s code figures come from a company that builds AI coding tools and says it uses Claude extensively. They cannot establish typical productivity or adoption at other employers. The METR trial was small and involved experienced developers working in familiar projects; its results may not predict outcomes for beginners, other types of work or newer tools.
METR’s later study also produced an unreliable signal. The group said some developers were reluctant to accept tasks without AI, and its raw estimates—18% less time for developers from the first trial and 4% less for newly recruited developers—had confidence intervals that included no effect. METR said the results suggested developers were likely more sped up in early 2026 than in early 2025, but offered only weak evidence about how much.
The job figures do not isolate AI’s effect from other forces shaping hiring. Nor does one learning experiment settle how AI affects education over time, with different prompts, teaching methods or levels of supervision. The available material does not establish a single best curriculum or show what the long-term career outcomes will be for people learning programming now.
What Learners and Employers Should Track
The next useful evidence will come from larger and repeated studies that compare current AI tools with unaided work across different experience levels and real-world projects. Those studies need to measure more than completion time: code quality, bugs, security, maintenance and what participants retain are also relevant outcomes. The source material does not identify a scheduled date for such results.
Labor-market tracking will show whether the recent rise in software postings continues and whether entry-level roles recover relative to senior openings. For learners, the evidence supports practicing both AI-assisted work and unaided fundamentals, then checking whether they can explain, test and debug what a tool produces. That is a practical response to the findings, not a guarantee of employment or a settled prescription.
Key Questions
Does the evidence show that learning to code is a waste of time?
No. The evidence shows that AI can generate substantial code in some settings, but it does not establish that programming knowledge is obsolete. Requirements, testing, debugging and security judgment remain part of the work described in the source material.
Do AI tools make developers faster?
Results differ by study. METR’s 2025 trial found experienced developers took 19% longer with the tested AI tools, while a 2022 Copilot experiment found participants completed a small coding task 55.8% faster. The tasks and tools differed, and METR said its later evidence on speed gains was weak.
What did the study of junior developers find?
In an Anthropic experiment about learning a new library, mostly junior developers who used AI scored 50% on a follow-up quiz, compared with 67% among those who coded by hand. The largest difference concerned debugging. The result applies to that experiment and does not settle how all AI-assisted learning works.
Is entry-level software hiring disappearing?
The cited Indeed data show entry-level jobs were 4.5% of US software development postings in the first quarter of 2026, while senior roles were 69.3%. But postings were also about 23% higher year over year in November 2026. These figures do not prove AI caused the entry-level share or establish that hiring has disappeared.
Source: ThorstenMeyerAI.com
